fix(publish): Change prepare to prepublishOnly

[spencerwilson-optimizely]

Summary

• Convert prepare hook to a (not-equivalent) prepublishOnly, reducing the number of times unit tests are run in prepare-literate builds by 1.

#129 introduced a regression where unit tests stopped being run for Node.js v4 and v6. Prior to that PR, unit tests were run

• v4 and v6: once
• v8, v9, v10, ... (versions whose npm understands the prepare hook): twice

After that PR was merged, the number of times unit tests were run in each build went down by 1, for every build. That means v4 and v6 stopped running unit tests altogether.

2faa9c4 fixed the regression by unconditionally running unit tests on all matrix-based builds (and explicitly adding a build to run xbrowser tests). With this, Node.js 8+ builds resumed running unit tests twice (example).

Regarding the prepare hook: npm@4 added the prepare hook with that fires on prepublish and postinstall. This is strange. Why run tests when someone installs the package? They see the mocha output in their shell, which is likely to surprise.

This PR stops the unconventional test runs on postinstall.

Test plan

CI.

[coveralls]

Coverage decreased (-0.1%) to 97.11% when pulling 847726e on sw/postinstall into 2faa9c4 on master.

[tylerbrandt]

Hm I'm not sure what happened here; at the time that I added the "prepare" hook I didn't remove anything; I was just attempting to codify in code what we currently require from users to deploy. So the regression must have come later (or earlier).

[spencerwilson-optimizely]

@tylerbrandt I think this was a regression due to #129. Before that PR, we unconditionally ran unit tests once, and we ran them twice in the case of npms which understand prepare.

[tylerbrandt]

Ok, sounds good. Can you update the summary to remove the misleading stuff?

[spencerwilson-optimizely]

Was in the process of doing that, you're just darn fast at reviewing, haha.